question about using pretranied G2P model and train G2P model myself

[li-henan]

dear author， could I ask whether it is useful to train G2P model myself？after reading materials，my understanding is that g2p model training needs a dictionary firstly，but if I have a dictionary，why I straightly align my wav and txt. Is it right that dictionary can be complemented using a g2p model newly trained？ could you help to explain the benefit following these steps？ mfa g2p data g2pmodel out_dictionary mfa train_g2p out_dictionary g2pmodel

thank for replying very much.

[Jackson-Kang]

Dear @Eie1 , First of all, thank you for question.

As you pointed out, there are some unnecessary steps in the tutorial code. If you already have grapheme-to-phoneme(g2p) mapping dictionary, we don't need to train g2p module from scratch. However, if you do not have mapping dictionary, you can use (1) rule-based(algorithmically operating) g2p module or (2) pre-trained g2p module (or g2p module trained with your own data).

Please, keep in mind that dictionary generated from other dataset may have Out-Of-Vocabulary(OOV) words. In this case, using the dictionary can generate OOV tokens in the TextGrid alignment file.

In that sense, I strongly recommend that dictionary be generated using your own data. If you don't have enough data to train, then consider using other dataset together. In this case, your g2p mapping dictionary must cover any vocabulary in your grapheme corpus.

Thank you.